[Madraka]

What

Fixes #116.

Durable Object suites ran out of local ports (EADDRNOTAVAIL).

• Miniflare 4's dispatchFetch sets undici's reset on every request: it sends connection: close, so each call costs one new TCP connection and one local port in TIME_WAIT for 30 s.
• RemoteDoSqliteDriver (@cloudhal/storage/testing, createMiniflareDoSqlite) sends every SqlDriver call that way, and SqlBlobStore adds more.
• On macOS, core-ts's store.test.ts peaked at 16,344 TIME_WAIT sockets against 16,384 ephemeral ports, and 37 tests in its Durable Object tail failed with EADDRNOTAVAIL.
• Fix: the harness now sends calls straight to the entry socket that await mf.ready resolves to, over a node:http keep-alive Agent (4 sockets, 1 s idle close).
• A non-200 response or a connection closed mid-response rejects with a plain Error, so a broken harness can never satisfy toThrow(StorageError).
• No Miniflare patch, so the ADR-0012 pin stays. No change to the test Worker or to the public { driver(db?), dispose() } shape.
• Patch changeset for @cloudhal/storage.

Studio teardown flake: "window is not defined" (CI run 39959905949).

• Studio's Vitest has no globals, so Testing Library never registered its own afterEach(cleanup).
• Trees stayed mounted after a file ended. Base UI / Floating UI work then committed outside act, and React's passive flush ran after happy-dom had removed window.
• apps/studio/test/setup.ts now runs cleanup() and vi.useRealTimers() after every test, and drains three setImmediate rounds after each file.

Verification (local, sequential):

	Before	After
pnpm --filter @cloudhal/core-ts test	37 failed (EADDRNOTAVAIL), 337 s	988/988 twice in a row, 0 EADDRNOTAVAIL, ~350 s
Peak TIME_WAIT	16,344	~4,000 (the D1 suites' startup burst; DO-only run: 11)
@cloudhal/storage do-sqlite.test.ts	17/17	17/17
components.test.tsx × 10	flaky in CI	10/10, no unhandled errors
Studio	–	79 files, 1275 tests, 0 unhandled, 0 act() warnings

Also run: biome, typecheck for storage and Studio, size:check, design:lint.

Process: three independent reviewers checked the change, and two of their findings held up when a separate agent tried to refute them. Both are fixed in this PR: the mid-response close now settles the promise (reproduced with a small Node script), and harness HTTP failures are a plain Error.

Found along the way, not in this PR:

• packages/mcp-server/src/render/checks/content.ts:72 writes /\s/ inside a template literal, so the in-page empty-section check tests for the letter s, not whitespace. pnpm lint:fix "fixes" it to /s/.
• The published @cloudhal/storage/testing (and analytics, comments, schedules) ./testing exports point at files build-packages.ts skips.

Checklist

• pnpm check and cargo test --workspace pass locally: the affected packages' typecheck and tests, size:check and design:lint were run one at a time (the owner's machine freezes when suites run in parallel). Full pnpm check and cargo were not run; no Rust or contract code changed.
• Contract change? None.
• User-visible change in a published package? .changeset/do-sqlite-keepalive.md (@cloudhal/storage: patch)
• Commits are signed off (git commit -s, DCO)
• Docs / AGENTS.md updated if behaviour or commands changed: storage README

🤖 Generated with Claude Code